I am coming to believe that these shady generative AI spelling misadventures are actually not at all what we think they are.
Our primary method of communication in language is subject to core ambiguities at the semantic layer.
Certainly, the training data is formative but insofar as language itself is concerned, these disjointedly discontinuous misrepresentations are actually accurate representations of a semantic layer of language that is in some sense free-floating and unbounded by conventions of spelling because language’s core mechanisms are not oriented towards or for the veridical purposes to which we might apply them.
Failure to represent language is simultaneously the success to represent its own underlying, generative properties.
Among other things, this gestures towards the whole truth/fake news and disinformation conversation as itself being grounded in an unconstructive misrepresentation.
In a nutshell – communication communicates.
Everything else is interpretation and the endemic indeterminacy of anything other than the most trivial of narrative artefacts represents a primary transmission medium for the production of yet more language, more technology and more adversarially competitive sociotechnical confusion.
Daedelus Kite 